The implementation and performance analysis of AWMMF parallel algorithm on GPU

Abstract

For most practical image processing systems, people often make the tradeoff between image processing quality and algorithm time complexity, because the best algorithms cannot meet the real-time requirement. However, with the introduction and development of CUDA (Compute Unified Device Architecture), GPU (Graphics Processing Unit) is widely used to implement parallel algorithms, which can speed up the massive and complex computing procedures. We once proposed an algorithm called AWMMF (adaptive window multistage median filter) for image salt-and-pepper denoising. In order to ensure its real-time performance, we implement the corresponding parallel algorithm with CUDA on GPU, and analyze its performance of several optimization techniques. The results not only show that the parallel algorithm on GPU is significantly superior to the serial one on CPU, but also indicate that reasonable optimizations can have a huge performance improvement.